Shazeer: multi-query attention for incremental decoding

What changed

Shared key and value representations across query heads to reduce decoder state traffic and improve incremental inference efficiency.

Scientific scope

Shared key and value representations across query heads to reduce decoder state traffic and improve incremental inference efficiency. This changes the attention architecture and can affect quality; it is not a lossless kernel optimization or the later grouped-query generalization.

Evidence

The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. This changes the attention architecture and can affect quality; it is not a lossless kernel optimization or the later grouped-query generalization.

Contributors

• Source-listed authors and technical contributors · SCIENCE: Source-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities.

• Fast Transformer Decoding: One Write-Head is All You Need

Provisional WRS assessment, edition 2026-10-16.1. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.

Full scientific record · Browse contribution grades · Request a correction